Research map / Vision-language and multimodal / Multimodal LLMs
Reasoning-Guided Part-Level Visual Grounding via Reinforcement Learning
Kazi Sajeed Mehrab, Hani Alomari, Najibul Sarker, Zaber Ibn Abdul Hakim, Chia-Wei Tang, Anuj Karpatne, Chris Thomas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with part-aware GRPO and OP-HRG prompt instead of zero-shot prompting (ablation) | OP-HRG · InstructPart (test) | gIoU: 31.45 -> 75.56 | Improved |
| add hierarchical grounding (object-first, containment, object hint reward) (ablation) | OP-HRG · InstructPart (test) | gIoU: 73.77 -> 75.56 | Improved |
| add reflective refinement (self-critique and improvement reward) (ablation) | OP-HRG · InstructPart (test) | gIoU: 73.98 -> 75.56 | Improved |
| add OP-HRG structure and part-aware rewards over plain localization rewards (ablation) | OP-HRG · InstructPart (test) | gIoU: 70.32 -> 75.56 | Improved |
| adapt VisionReasoner with GRPO on part data only (ablation) | VisionReasoner · InstructPart (test) | gIoU: 59.38 -> 62.33 | Improved |
| swap SAM2 mask decoder for SAM3 (ablation) | OP-HRG · part benchmarks | gIoU: 75.56 -> 75.77 (InstructPart) | No clear effect |
| add active visual perception at 4B (ablation) | OP-HRG · part benchmarks | gIoU: PartImageNet 56.87 -> 55.98; InstructPart 75.56 -> 76.34 | Mixed |
| scale backbone from 4B to 8B with AP (ablation) | OP-HRG · PascalPart | gIoU: 39.04 -> 40.28 | Improved |
| evaluate with plain single-answer prompt instead of reflective prompt (ablation) | OP-HRG · InstructPart (test) | gIoU: 75.56 -> 75.40 | No clear effect |
| part-centric RL training | Qwen3-VL-Instruct-4B · RefCOCO/+/g average | accuracy (%): 89.7 -> 86.5 | Hurt |
| OP-HRG-trained 4B vs VisionReasoner 7B | OP-HRG · ReasonSeg | gIoU: 63.6 -> 69.6 | Improved |
| remove external baseline from improvement reward (ablation) | OP-HRG · training dynamics | initial-answer IoU | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- VisionReasoner-7B — VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning (ICLR 2026) · better on 5 of 5 shared results
- UniPixel-3B — UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning (NeurIPS 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Connecting the Dots: Training-Free Visual Grounding via Agentic Reasoning (AAAI 2026 · similarity 0.91)
- GROUNDHOG: Grounding Large Language Models to Holistic Segmentation (CVPR 2024 · similarity 0.90)
- No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers (ICLR 2026 · similarity 0.89)
- Enhancing Part-Level Point Grounding for Any Open-Source MLLMs (CVPR 2026 · similarity 0.89)
- Investigating Compositional Challenges in Vision-Language Models for Visual Grounding (CVPR 2024 · similarity 0.88)
- GeM-VG: Towards Generalized Multi-image Visual Grounding with Multimodal Large Language Models (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
